Download a) Hebbian learning

Labmodule: Learning Rules
Introduction: Essential to the concepts of learning and memory is the ability of neurons to somehow
change in order to accommodate learned information. While this could conceivably occur through a
number of non-mutually exclusive physiological mechanisms (e.g. the growth of new neurons, the
development of new axonal projections, or even a significant alteration in the resting potentials of a
group of neurons), one mechanism in particular, the alteration of the weights of synaptic connections, is
believed to predominate in the developed brain. Though the complex biochemical pathways underlying
this phenomena is still an area of active research, several theoretical models of how this may occur have
been proposed. Here, two such theories, which are of particular use to computational neuroscientists,
will be examined in the context of the famous experiments performed by Ivan Pavlov during the turn of
the twentieth century.
While performing research on canine digestion, Pavlov noticed that his dogs would naturally salivate in
anticipation of a meal. Furthermore, if he, for several days in a row, presented food while also
sounding a tone, he showed that his dogs could eventually be trained, or “conditioned”, to salivate in
response to the tone alone. These two models will use a very simplified network, made up of only
seven neurons, to illustrate how the learning rules can mediate these behavioral observations. Activity
in the olfactory sensory neuron represents detection of the unconditioned stimulus (unconditioned
because the dog naturally salivates in response to the smell of food), whereas activity in the auditory
hair cells represents detection of the stimulus which is to be conditioned, in this case a bell. Activity
within the cortical neuron represents the successfully elicitation of a response, which in this case is
salivation.
Five auditory hair cells, abbreviated HC X, each respond selectively to sound of different frequency. The
peak of each cell’s tuning curve is centered at (100)(X) Hertz. The width of these tuning curves can be
adjusted by altering the discriminatory ability in the parameter panel. Each of these hair cells synapses
upon the cortical neuron controlling salivation. Initially these synaptic weights are set to zero, indicating
that the sound of a bell does not naturally cause the dog to salivate. One olfactory sensory neuron,
abbreviated OSN, responds whenever food is presented. The synapse between the OSN and the cortical
neuron has the maximum value, because the smell of food will always cause the dog to salivate.
Part 1: Hebbian Learning
1) Load the BioNB330 Software Program.
2) Click on “Tutorial 6: Learning Rules” in the Main Menu.
3) Read the introduction, and proceed to the first model.
4) The first paradigm employed is the basic Hebbian learning rule. Qualitatively, this states that when
two neurons are active at the same time, the synapse between them will be strengthened. This rule
does not take into consideration which neuron spiked first, nor does it provide a strict temporal
definition for “at the same time.”
Equation:
The Hebbian Learning Rule:
Wij    X i  X j
Where μ is defined as the rate of learning, Xi is the activity of postsynaptic neuron i, and Xj is the activity
of presynaptic neuron j.
Task 1) Set  to 1x10-4 , present the bell from 0 to 50 ms, the food for 20 to 50 ms. Set trial
number to 10 and find out how many trials of presentations are needed for the cortical neuron to
respond to the bell alone. After the first time this happens, can you still observe a change? Why?
Create a plot showing learning rate versus number of trials needed for the cortical neuron to
respond to food alone. What other parameters could you change that affect the number of trials
needed? How could you affect learning rate in real life – i.e. when training an animal?
Task 2) Now reduce the discriminatory ability of the hair cells to 0.6 (this effectively broadens each of
their tuning curves). Now how many trials does it take to condition the cortical neuron? Again, if you
noticed a difference, why? Why are the Hair cells not all responding to the tone with the same spike
latency? After conditioning, DO NOT RESET but set food presentation to zero, set learning rate to zero,
number of trials to one and now systematically change the bell frequency in 20 Hz increments. Draw a
graph showing the frequency-response curve of the cortical neuron. If it responds to frequencies other
than 300 Hz, explain why. How could this be prevented?
Questions:
1. In what situations would it be advantageous for a system, by default, to maintain a very high rate of
learning (large value of μ)? When would it be better for a system to keep μ at a relatively low value?
2. Physiologically speaking, what are some ways in which an individual neuron might be able to alter it’s
own susceptibility to synaptic modification? How might a larger network accomplish the same goal?